Re: wait_on_page_bit_common(TASK_KILLABLE, EXCLUSIVE) can miss wakeup?

[Nicholas Piggin]

Excerpts from Linus Torvalds's message of June 30, 2020 2:36 am:
> On Mon, Jun 29, 2020 at 6:16 AM Nicholas Piggin <npiggin@xxxxxxxxx> wrote:
>>
>> No, ignore this part (which you explained well it was just a thinko,
>> and your patch of course would not have worked if this was the case):
>> the exclusive wake up doesn't get lost if schedule() was called because
>> state goes back to running regardless of what woke it.
> 
> Right.
> 
> The normal pattern for a wait-loop is
> 
>  - add yourself to the wait-queue and set yourself "sleeping" with a
> memory barrier.
> 
>  - test for the condition, exit if ok
> 
>  - go to sleep
> 
> and that pattern doesn't have the race.
> 
> The other common pattern is to check for the "do I need to sleep at
> all" at the *top* of the function, long before you bother with any
> wait-queues at all. This code does that odd "let's check in the middle
> if we need to sleep" instead, which then caused the bug.
> 
> So we had an odd setup because of three different wait conditions that
> had different rules for what they could/should do before sleeping, and
> then not sleeping reliably at all.
> 
> We could also fix it by just splitting out the three cases into their
> own wait routines that match the normal pattern. The bug really
> happened because that wait-loop is doing things such an odd way due to
> all the different cases..
> 
> I actually think it would be a lot more readable if it was three
> different cases instead of trying to be one "common" routine.
> 
> The *common* parts is the special PG_locked logic at the top, and the
> thrashing/delayacct code at the bottom.
> 
> And *that* could be a true common helper, but the wait loop (which
> isn't even a loop for the DROP case) are fundamentally different and
> probably should be separate functions.
> 
> So I think my "one-liner" fixes the problem, but I'd certainly be open
> to somebody cleaning this up properly.

I think it does. I would rather my patch which doesn't add a barrier to 
the "fast" path though (by the time we get here I think the fast path is
that we are sleeping on the page, doing the IO).

Alternatively (although this still adds more branching than necessary at
least it doesn't have a barrier for the sleeping case), you could set
the running state as an else case if you don't schedule.

Thanks,
Nick